Embodied AI shifts training into real production workplaces
PINGWEST reports that embodied AI training is moving from exhibition halls into real workstations, with June 2026 action by China’s MIIT and SASAC promoting production-site training for humanoid robots and embodied AI. The report cites an estimate that universal embodied models require at least tens of millions of hours of real interaction data, while only millions of hours of high-quality robot data are available globally. It highlights factory deployments, feedback pipelines and the LWD framework for returning deployed interaction data to policy training.
The report frames real-world training sites as infrastructure for linking robot data collection, annotation, model training, deployment and failure feedback within one operational loop.
机器人在哪里练,练成后用在哪里?越来越多的探索指向同一个答案:把训练放进真实的工位,这就是“实景实训平台”。据中国信息通信研究院《具身智能训练场研究报告(2026年)》及多家机构测算,训练通用具身模型至少需要千万小时级真实交互数据,而目前全球可用的高质量真机具身语料仅在百万小时量级,相差约一个数量级。若主要依靠专家遥控真机采集,产出往往不过每天数十小时。缺的不是又一间展厅,而是能把采集、标注、训练、部署失败样本再送回去的数据流水线。
在张江的觅蜂科技实训实验室,00后数据采集组长朱绵锐正与一台机器人并肩而立。“哪怕只是拿瓶子、放物件这样简单的小动作,我们也需要反复‘教学’几百遍,它们才能慢慢熟练。”朱绵锐如是说。这间被誉为“机器人学校”的实验室,折射出当前“语料饥渴”的现实困境。大语言模型可依托互联网海量文本实现“自学”,而机器人只能在真实世界中一寸寸“磨”出动作数据。
#China Academy of Information and Communications Technology#Ministry of Industry and Information Technology#State-owned Assets Supervision and Administration Commission#AGIBOT#Shanghai Electric#SAIC General Motors#Figure AI#BMW#Apptronik#Mercedes-Benz#Huawei#LWD